---

Requirements and Installation

• PyTorch version >= 1.5.0
• Python version >= 3.6
• For training new models, you'll also need an NVIDIA GPU and NCCL
• To install this version of fairseq go to repository root level and run:

pip install --editable ./

# on MacOS:
# CFLAGS="-stdlib=libc++" pip install --editable ./

In order to repeat results from the Reinforcement Learning for on-line Sequence Transformation paper:
Download and preprocess the data:

# Download and prepare the data
cd examples/translation/
bash prepare-iwslt14.sh
cd ../..

# Preprocess/binarize the data
TEXT=examples/translation/iwslt14.tokenized.de-en

fairseq-preprocess --source-lang de --target-lang en \
    --trainpref $TEXT/train --validpref $TEXT/valid --testpref $TEXT/test \
    --destdir data-bin/iwslt14.tokenized.de-en \
    --workers 20

This preprocesses data for German-English translation. To train and evaluate English-German translation models, switch --source-lang and --target-lang values in fairseq-preprocess command.

Training RLST on WMT15:

CUDA_VISIBLE_DEVICES=0,1,2,3 fairseq-train data-bin/wmt15_de_en --arch rlst --criterion rlst_criterion --no-epoch-checkpoints \
--eval-bleu --eval-bleu-detok moses --eval-bleu-remove-bpe --best-checkpoint-metric bleu --eval-bleu-args '{"beam": 1}' --maximize-best-checkpoint-metric \
--rnn-hid-dim 512 --rnn-num-layers 2 --rnn-dropout 0.0 --src-embed-dim 256 --trg-embed-dim 256  --embedding-dropout 0.0 \
--max-tokens 4096 --max-epoch 100 --optimizer adam --clip-norm 10.0 --lr 1e-3  --weight-decay 1e-5 --left-pad-source --rho 0.99 \
--epsilon-min 0.2 --epsilon-max 0.2 --rtf-delta 1.0 --N 200000 --m 7.0 --discount 0.90 --eta-min 0.02 --eta-max 0.2 \
--save-dir checkpoints/rlst

Training the transformer model:

CUDA_VISIBLE_DEVICES=0 fairseq-train data-bin/iwslt14.tokenized.de-en --arch transformer_iwslt_de_en \
--optimizer adam --adam-betas '(0.9, 0.98)' --clip-norm 10.0 --lr 5e-4 --lr-scheduler inverse_sqrt --warmup-updates 4000 \ 
--dropout 0.3 --weight-decay 0.0001 --max-tokens 4096 --eval-bleu --eval-bleu-args '{"beam": 1}' --eval-bleu-detok moses \ 
--eval-bleu-remove-bpe --best-checkpoint-metric bleu --maximize-best-checkpoint-metric --no-epoch-checkpoints \ 
--max-epoch 100 --save-dir checkpoints/transformer

Training the encoder decoder LSTM model:

CUDA_VISIBLE_DEVICES=0 fairseq-train data-bin/iwslt14.tokenized.de-en --optimizer adam --lr 1e-3 --clip-norm 10.0  --max-tokens 4096 \
--save-dir checkpoints/lstm/ --arch lstm_wiseman_iwslt_de_en --eval-bleu --eval-bleu-detok moses --eval-bleu-remove-bpe \
--eval-bleu-args '{"beam": 1}' --best-checkpoint-metric bleu --maximize-best-checkpoint-metric --no-epoch-checkpoints \ 
--max-epoch 100

Trained LSTM and transformer models can be evaluated on the test set using the fairseq-generate command:

CUDA_VISIBLE_DEVICES=0 fairseq-generate data-bin/iwslt14.tokenized.de-en/ --path <path to model checkpoints directory>/checkpoint_best.pt --beam 1 --remove-bpe --quiet

To test RLST you also need to provide --left-pad-source flag:

CUDA_VISIBLE_DEVICES=0 fairseq-generate data-bin/iwslt14.tokenized.de-en/ --path <path to model checkpoints directory>/checkpoint_best.pt --beam 1 --remove-bpe --quiet --left-pad-source